Show HN: Open-source implementation of Stanford's self-learning agent framework (github.com)

Kayba released an open-source implementation of Stanford/SambaNova’s Agentic Context Engine (ACE), a framework that makes LLM agents "self-improving" by learning from execution feedback instead of fine-tuning. The repo (pip-installable as ace-framework) provides a plug-and-play library that works with 100+ LLM providers via a LiteLLM client (OpenAI, Anthropic, Google, Ollama, etc.), quickstart examples, LangChain support, and demonstrated gains of 20–35% on complex tasks. A live demo (the “seahorse emoji” hallucination) shows ACE reflecting on mistakes and updating strategies to avoid repeat errors. Technically, ACE splits agent behavior into three cooperating roles—Generator (executes tasks using current strategies), Reflector (analyzes outcomes), and Curator (creates incremental updates)—with a Playbook serving as a living, transparent repository of learned tactics. Learning happens in-context via incremental playbook updates (no training data or parameter fine-tuning), which the project claims avoids context collapse and preserves edge-case knowledge and tool-usage patterns. For researchers and engineers this means reproducible, interpretable continual learning for agents, faster iteration on multi-step workflows, better tool selection, and safer, more reliable behavior without the compute cost of model retraining.